Connecting an MCP Server as a Workflow Tool
Workflows can now call out to any MCP server as a step-level tool — a separate, opposite-direction feature from the platform's own public, read-only MCP server.
Workflows now supports Model Context Protocol (MCP) servers as a step type, so a workflow can call out to any MCP server as a tool — not just the capabilities we've built in ourselves.
What this adds
You can register an MCP server's connection details from a dedicated panel in the editor, and from there a step can be configured to call any tool that server exposes. This is managed like any other connection — list your registered servers, add new ones, remove ones you no longer need — and it opens a workflow up to whatever an MCP server can do, without us having to build a bespoke integration for each one.
Don't confuse this with our own MCP server
It's worth being precise here, because the platform actually has two separate MCP integrations pointed in opposite directions. This feature — an MCP server as a workflow step — is inbound: your workflow is the client, calling out to someone else's server. Separately, the platform also runs its own public MCP server that exposes a small, deliberately read-only set of tools back out to other MCP clients — nothing in it can edit a workflow definition. The two aren't related beyond sharing the same protocol; one lets your workflows use other people's tools, the other lets other tools read (never write) yours.
Why now
MCP has moved fast this year as a way for AI tools to interoperate, and treating any MCP server as a first-class step type — the same way an HTTP request or an AI agent step works — means a workflow's toolset isn't limited to what we've pre-built.
If you're already running an MCP server for another part of your stack, you can point a workflow at it directly today.